Published in:
Open Access
01-12-2015 | Research article
County-level hurricane exposure and birth rates: application of difference-in-differences analysis for confounding control
Authors:
Shannon C. Grabich, Whitney R. Robinson, Stephanie M. Engel, Charles E. Konrad, David B. Richardson, Jennifer A. Horney
Published in:
Emerging Themes in Epidemiology
|
Issue 1/2015
Login to get access
Abstract
Background
Epidemiological analyses of aggregated data are often used to evaluate theoretical health effects of natural disasters. Such analyses are susceptible
to confounding by unmeasured differences between the exposed and unexposed populations. To demonstrate the difference-in-difference method our population included all recorded Florida live births that reached 20 weeks gestation and conceived after the first hurricane of 2004 or in 2003 (when no hurricanes made landfall). Hurricane exposure was categorized using ≥74 mile per hour hurricane wind speed as well as a 60 km spatial buffer based on weather data from the National Oceanic and Atmospheric Administration. The effect of exposure was quantified as live birth rate differences and 95 % confidence intervals [RD (95 % CI)]. To illustrate sensitivity of the results, the difference-in-differences estimates were compared to general linear models adjusted for census-level covariates. This analysis demonstrates difference-in-differences as a method to control for time-invariant confounders investigating hurricane exposure on live birth rates.
Results
Difference-in-differences analysis yielded consistently null associations across exposure metrics and hurricanes for the post hurricane rate difference between exposed and unexposed areas (e.g., Hurricane Ivan for 60 km spatial buffer [−0.02 births/1000 individuals (−0.51, 0.47)]. In contrast, general linear models suggested a positive association between hurricane exposure and birth rate [Hurricane Ivan for 60 km spatial buffer (2.80 births/1000 individuals (1.94, 3.67)] but not all models.
Conclusions
Ecological studies of associations between environmental exposures and health are susceptible to confounding due to unmeasured population attributes. Here we demonstrate an accessible method of control for time-invariant confounders for future research.